schema_version string | record_kind string | machine dict | model string | artifact_id string | artifact_revision string | artifact_format string | quantization string | runtime string | context_tokens int64 | measurement dict | verification_method string | observed_at timestamp[s] | note string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0.9 | riggraph_measurement | {
"os": "Darwin",
"arch": "arm64",
"cpu": "Apple M4 Max",
"ram_gib": 48
} | mlx-community/Qwen3-0.6B-4bit | hf://mlx-community/Qwen3-0.6B-4bit/mlx | 73e3e38d981303bc594367cd910ea6eb48349da8 | MLX | 4-bit | mlx-lm | null | {
"generation_tps": 560.85,
"prompt_eval_tps": 1629.77,
"total_latency_s": 0.2639
} | null | 2026-09-24T17:14:46 | Initial v0.9 live smoke; context was not persisted by the pre-hardening receipt schema. |
0.9 | riggraph_measurement | {
"os": "Darwin",
"arch": "arm64",
"cpu": "Apple M4 Max",
"ram_gib": 48
} | mlx-community/Qwen3-0.6B-4bit | hf://mlx-community/Qwen3-0.6B-4bit/mlx | 73e3e38d981303bc594367cd910ea6eb48349da8 | MLX | 4-bit | mlx-lm | 4,096 | {
"generation_tps": 560.23,
"prompt_eval_tps": 1687.91,
"total_latency_s": 0.263
} | race-v2 | 2026-09-24T17:31:26 | Verification-only recheck of the exact pinned artifact from local cache; no acquisition/load mutation. |
LLMRig Benchmark Passports
Privacy-safe, machine-specific execution evidence generated by LLMRig.
This dataset is deliberately small in v0.9. Its purpose is to demonstrate the Benchmark Passport / RigGraph evidence schema, not to claim a universal leaderboard.
What a passport represents
Each row describes one observed combination of:
machine × model × artifact × quantization × runtime × context × measurement
Missing facts remain missing. LLMRig does not convert unknown values into synthetic values.
Current evidence
The initial records come from the v0.9 Apple Silicon release smoke:
- Apple M4 Max
- 48 GiB unified memory
mlx-community/Qwen3-0.6B-4bit- MLX / 4-bit
- MLX-LM
- exact Hugging Face revision
73e3e38d981303bc594367cd910ea6eb48349da8
The second row is the verification-only recheck after v0.9's receipt/context hardening. It re-used the exact pinned artifact from local cache and performed no acquisition/load mutation.
Important limitations
- Throughput is hardware-, runtime-, model-, context-, and workload-specific.
- Performance measurements do not establish model quality.
- The first record predates receipt-context hardening and therefore keeps
context_tokensasnull; it is intentionally not backfilled. - Private filesystem locators, secrets, account identifiers, and tokens are excluded.
Files
benchmark_passports.jsonl — one JSON object per measured observation.
Reproduce locally
pip install llmrig==0.9.0
llmrig run mlx-community/Qwen3-0.6B-4bit
llmrig verify
The exact values you observe on your machine may differ.
- Downloads last month
- 21